Computation of hold blocks was done by storing ranges (with start and
an end, and likely overlapping) in a tree keyed only by the block start.
This cannot work well, and there even were comments that it is not
reliable in complex cases.

A much better way to deal with it is to split all ranges so they don't
overlap. The most thorough way of doing this is to split at all and every
known keyframe, and in this case the data can actually be stored in the
key column data structures, avoiding the need for a second tree.

In practice, splitting requires a pass to copy this data to newly added
keys, and the necessity to loop over all keyframes in the range being
added. Both are linear and don't add excess algorithmic complexity.